| Author | Message | ||
     Erik Skibsted |
Hi I am helping out a master student with his NIR project. He is measuring the filtered from samples removed from a pharmaceutical fermentation process. We started with key metabolites i.e. glucose. We then started to look on the pharmaceutical product itself i.e. a protein. We use a Bruker, MPA, with heat controlled 1 mm pathlength vials, using transmission measurements. One of problems we are strugling with is what background is the best. We use two approaches (1) air as background (2) the vial with a buffer solution as background. In my oppinion the application is difficult. The protein concentration is low and water is the major component. What is you experiences, what background is the most promising (stable) is such an application. Also what about preprocessing methods? Can OSC add something? Second. I am working on the calibration models, and I am getting the impression that when the PLS models for the protein looks promising, it is a kind of 'in-direct' calibration using the glucose in the samples to make a calibration with, because normally in the fermentation there is an inverse relationship between the protein and glucose i.e. glucose is used as feed for the microorganisms in order to produce the protein. All comments are welcomed Erik Skibsted | ||
| Bruce H. Campbell (Campclan) |
Erik, I would be reluctant to use a vial/cuvette filled with the buffer solution. My reason is that the absorbance due to water in the buffer solution will be less than the samples. The amount of decrease will be related to the amount of sample. Put in another way, the sample dilutes the water. By instrumentally substracting water, you may be introducting an error. If water is not of any interest whatever, you may want to use wavelength selection to remove the water bands from the calibration. However, the very decrease of absorbance of the water bands should relate to the concentration of the sample, which may be useful by itself. Can OSC add something? Try it with and without OSC to find the better calibration. If you do not have a regression "band" in the correct position for an NH band, you probably are not detecting the protein at all. Indirect calibrations are acceptable, in my opinion, but do require much more care and calibration samples to ensure the ruggedness of the calibration. Bruce | ||
| Erik Skibsted |
Thank you for your comments Bruce. I am also myself more in favor of using air as background, also I think the spectroscopically interpretation os much easier. In all the papers about NIR calibration of protein in water solutions I recall that they usually operate with large datasets and 5-9 PLS components. Because of the very complex nature of the background variation i.e. biological salts, metabolits etc. I can accept that. So what you say is that look on the 'spectral signature' of the final PLS regression vector to see if there are peaks at the positions where you would expect NH band absorption?? Erik | ||
| Gabi Levin |
Hi Erik, Bruce is absolutely right. The water should stay, the PLS1 can make use of it. I trust that you use absorption files rather than first derivative. Later it will become clear why it is preferrable at this stage to use absorption spectra rather than first derivative. I had myself tough time, in some cases "no deal" situation, meaning I refused to continue to waste the potential customer and my time. When you say concentration is low, how low? In addition to the advice from Bruce, a couple of points: 1. Proteins in solutions are treacherous - the hydrogen bonding and the spatial arrangements of the protein may vary with other compositional changes, in pH, in cations, in anions, etc. The result of such changes can be real problematic, because the spectrum will change, but the concentration determined by reference method remain conatant. 2. When looking for evidence of relation between the calibration and the concentration of the determined species, we need to look at something I am used to from Unscrambler - the loading weights for each PC. In Unscrambler this plot indicates in what wavelength the PC finds correlation to the concentration of the species. Altough the evidence may not appear in the first or second PC's it should start to show up at least in the third and up. In many cases I find a very strong correlation between the spectral features of the species and the loading weights straight from the first PC. You should also, if you have Unscrambler, look for each PC how much of the Y-data it explains. If the PC's where you find strong relation to the spectrum of the species in question explain large % of the Y-data it enhances the reliability of the calibration. 3. If it so happens that the glucose and the protein are having an inverse relation, the degree of caution should be even higher. In that case what is the range of glucose in the samples? Is it possible to create a set where the glucose will vary in an arbitrary way? 4. If not, another way to determine if the calibration is using the inverse ratio is by looking at the loading weights of the glucose and the protein. If the loading weights for the PC's happen to be mirror image, or near mirror image of each other then you are in trouble. It is a tedious process, and I do not know what PLS1 software you use and if you have these options, but if your analyzer allows you to convert your spectra into Unscramler format (actually you can be lucky if you can convert to Excel and copy from Excel into the Unscrambler)you can run regresions with the Unscarmbler (you can down load a trial package from the CAMO web site) and then you can have all these loading weights etc. If you need help with Unsrambler, let me know and I will try to help. I hope this was of any use to you. Gabi Levin Brimrose | ||
| David W. Hopkins (Dhopkins) |
Erik, I agree with Bruce and Gabi that the reference should be air. It will be much more stable than using vials of buffer, which will also be a source of variability. However, you may also wish to try first or second derivatives. They can remove some of the sources of variance and make models that require fewer factors. They could force the regression to use more of the chemical information instead of the indirect information. It will take only some more time to evaluate these pretreatments. I find second derivatives a little easier to interpret than first, but first derivatives are not so bad. I find that the best way is to compare the derivatives of the pure components with the loadings or the B-vector of the models, so you can make the direct comparison of bands. Sounds like an interesting project. I hope you let your student do some of the work, so he/she can have the fun of discovery too. Regards, Dave | ||
| Erik Skibsted |
Hi All thanks for your comments. I´n my mind I am also in favor of air (because of less chance of variability) I use MatLab and the PLS toolbox 2.1 so I think I have a lot of possibilities of inspecting the various PLS loadings. I know from the litterature that generally very complex models are expected using a lot of samples. I have a feeling that these types of data could benefit from neural network models? instead of PLS1 models Erik | ||
| Gabi Levin |
Hi, David you are right about derivatives, however, they also add noise, so the advantage should be weighed in actual testing the quality of predictions. My experience with liquids (clear ones at least) that 2nd derivative does not contribute that much. The reason I suggest absorption at least for the evaluation of the regression and on what it is based is that it is easier to correlate to the spectrum of the materials. Problem with the proteins is that Erik probably doesn't have a solution of pure protein. If he does, that could be a huge advantage in determination of the reliability of the calibration. Thanks, Gabi |